ISSN (Online) : 2319 - 8753
ISSN (Print) : 2347 - 6710
International Journal of Innovative Research in Science, Engineering and Technology
Volume 3, Special Issue 3, March 2014
2014 International Conference on Innovations in Engineering and Technology (ICIET’14)
On 21st & 22nd March Organized by
K.L.N. College of Engineering, Madurai, Tamil Nadu, India
Automatic Segmentation of Pulmonary Lobe
Using Marker Based Watershed Algorithm
Ulaganathan Gurunathan#1, Brindha Manohar*2, Banumathi Arumugam#3, Vijayakumari Pushparaj#4,
Balachandar#5
#1
Department of Oral & Maxillofacial surgery, Best Dental Science College, Madurai, India.
*2
Department of ECE, Thiagarajar college of Engineering, Madurai, India.
#3
Department of ECE, Thiagarajar College of Engineering, Madurai, India.
#4
Department of ECE, Thiagarajar college of Engineering, Madurai, India.
#5
Department of ECE, Thiagarajar College of Engineering, Madurai, India.
ABSTRACT— Segmentation is an important process in
the field of medical imaging, as it can provide detailed
information of an image. In this work, segmentation of
pulmonary lobe is carried out which is useful for the
clinical interpretation of CT images, to access the early
presence and the characterization of several lung diseases.
This segmentation process is challenging for severely
diseased lung or lung with incomplete fissures. Existing
methods highly rely on the detection of fissures whereas,
this technique becomes less reliable in cases of
abnormalities. In order to overcome this, automatic
segmentation of the lung lobe is performed using marker
based watershed algorithm. The proposed algorithm
consists of five stages. The first stage is the segmentation
of blood vessel using thresholding. The second step
involves the segmentation of fissures based on the Eigen
values of the Hessian matrix. The third stage is the
segmentation of bronchial tree through region growing.
Fusing these three results to form the cost image is the
fourth step. The final stage is the application of watershed
algorithm for the cost image. The run time, mean and
accuracy calculation were performed and the results
showed that the proposed method possess minimum
runtime, mean distance when compared with the existing
methodology.
KEYWORDS— fissure, lung lobe,
watershed transform, hessian matrix
cost
I. INTRODUCTION
The human lungs possess five lobes which are separated
by visceral pleura termed as pulmonary fissure. The three
lobes in the right lung namely, right upper, right middle
and right lower are separated by right minor fissure and
right major fissure respectively. The two lobes in the left
lung namely, left upper and left lower are separated by
left major fissure. Fig.1 represents the structure of lung.
Lung lobe segmentation is important in clinical
applications especially for diagnosis and treatment of lung
diseases. The location and distribution of pulmonary
diseases are important parameters for the selection of a
suitable treatment. Locally distributed emphysema can be
treated more effectively by lobar volume resection than
homogeneously
distributed
emphysema.
Another
application is quantitative monitoring of pulmonary
diseases such as emphysema or fibrosis. Progression of
the diseases can be determined accurately through the
lobe wise analysis.
image,
Fig.1 Structure of lung
Copyright to IJIRSET
www.ijirset.com
M.R. Thansekhar and N. Balaji (Eds.): ICIET’14
1182
Automatic Segmentation of Pulmonary Lobe using Marker Based Watershed Algorithm
The visualization of entire lung within few second is
possible through Computed Tomography (CT). Manual
segmentation is highly time consuming, since 400 slices
with sub-millimeter resolution in each direction is taken in
typical scans with high anatomical details. Thus there is a
great requirement for automatic segmentation of lung
lobe. Due to the variations in anatomical structure of the
lobes and incomplete fissure, segmentation of lung lobes
is difficult. There are two main reasons where the fissure
becomes incomplete or unrecognizable, they are:
pathologies can deform the lobe and thus the fissure
becomes unrecognizable and due to normal lung
parenchyma fissures are often not complete. Radiologists,
in such cases acquire information from the blood vessel
and bronchial tree to infer the lobar boundaries.
functions representing the fissures, the lungs were divided
into five lobes. No anatomical information of bronchi or
vessels were taken into account in cases of incomplete
fissures. Segmentation took on an average 25 min for one
case whose slice thickness is around 0.625–1.25mm.
This work proposes an automatic lung lobe
segmentation method that uses information from
automatic segmentations of the bronchi, vessels, and
visible fissures in a 3-D watershed transformation to be
both robust against incomplete fissures and accurate at
visible fissures. Although all previously published lobe
segmentation methods described above were evaluated, it
is not possible to compare the results directly because
evaluation was performed on different datasets and with
different evaluation measures (volume overlap, distance
to the fissures, visual inspection).
The literature shows that, there are number of methods
proposed for the lung lobe segmentation. Van Rikxoort et
al., presented a method for lobar fissure segmentation [1]
II. METHODOLOGY
in CT data. For patients with complete pulmonary
fissures, a segmentation of these fissures is sufficient to
A lobe segmentation method is developed which
obtain lung lobe segmentation. Since in many cases combines anatomical information from the lungs, vessels,
fissures are incomplete, additional processing steps are airways, and lobar fissures [6] to obtain the lobes using a
required to obtain lobe segmentation. Nevertheless, the watershed-based segmentation method. Fig.2 provides an
segmentation of visible fissures offers a good basis for overview of the segmentation process. The method starts
lobe segmentation and can also be used for other purposes by segmenting the lungs, vessels, airways, and fissures,
with clinical relevance such as quantifying the which are later combined into one cost image for the
completeness of the fissures, which is an important feature watershed segmentation process.
for treatment planning of patients with emphysema. In [2],
A. Preliminary Segmentations
pulmonary lobes and segments were found by supervised
classification. The fissures were enhanced and segmented
1) Pulmonary Vessels: Based on the assumption that
by the eigenvalues and eigenvectors of the Hessian there are usually no major vessels at the lobar boundaries,
matrix. Subsequently, features such as the position the distance to the pulmonary vasculature is a suitable
relative to the fissures provided a labeling to a pulmonary feature to detect lobar boundaries. To quantify the
lobe for every voxel inside the lung. The second lobe absence of vessels at the lobar boundaries, a coarse
segmentation method by van Rikxoort et al., is an segmentation of the pulmonary vasculature is sufficient.
automatic multi-atlas approach [3]. First, the lungs, There is high contrast between blood vessels and lung
fissures, and bronchi were segmented automatically and parenchyma that enables a coarse segmentation of the
combined into one cost image. A fast registration of this pulmonary blood vessels by thresholding the data inside
result with a set of five atlases with complete fissures gave the lung region. The goal is to include as many vessels as
the best matching atlas that was chosen for a fine possible but exclude fissures and other dense structures.
registration to get the lobe segmentation result. Atlas- Before thresholding, a downscaling with clamping is
based methods are generally time consuming, applied to reduce memory requirements. Using (1) the
segmentation of one case took 2 hours on an average. dataset is scaled down to the 8-bit range, where L is the
Another disadvantage of this approach was that scans with lung mask v is the image pixel.
lobar shapes not represented in the atlas set were unlikely
to be segmented correctly. Ukil and Reinhardt presented Vds =
max (0, min(254,(vorig+1024)/4)), v L
(1)
pulmonary lobe segmentation [4]. In the first step, the
255,
otherwise
lobes were segmented by a watershed transformation
based on a distance map of the vasculature and markers The resulting dataset is thresholded to receive the vesselfrom the labeled bronchi tree. In the second step, 3-D mask V
optimal surface detection was performed in a region of
interest (ROI) around the initial segmented fissures to
V = 130 ≤ Vds ≤ 255
(2)
refine the lobe boundaries. As a last step, incomplete
fissures were extrapolated based on a fast-marching
As reported by Bianca lassen, the intensity values of
method.
Pu.et.al. proposed an automatic lobe pulmonary vessels lie within 130 to 255, [7]. Thus fixed
segmentation [5] method that started by detecting plane threshold of 130 was empirically estimated on an
patches in sub volumes in the lungs. From these patches independent dataset and proved to be a good trade-off
the pulmonary fissures were inter- and extrapolated using between sensitivity and specificity.
implicit radial basis functions. Based on the implicit
Copyright to IJIRSET
www.ijirset.com
1183
M.R. Thansekhar and N. Balaji (Eds.): ICIET’14
Automatic Segmentation of Pulmonary Lobe using Marker Based Watershed Algorithm
2) Pulmonary Fissures: The first step of the fissure
segmentation process is an enhancement of the fissures
based on the eigenvalues of the Hessian matrix that gives
a fissure probability for each voxel [7,8].
Fig. 2 Pipeline of the proposed method
The relation between the eigenvalues of the Hessian
matrix describes the local image structure. The proposed
fissure enhancement approach characterizes fissure voxels
as: 0˂˂ ||˂  and || ≈ 0, where δ describes the ||
value for vessels.  is introduced to discriminate between
fissures and vessels since vessels usually exhibit a larger
|| compared to fissures because of their stronger image
contrast. From these characteristics two features are
derived using (3) and (4).
Fstructure = exp [-(()^6) / (^]
Fsheet = exp [-(() ^6)/ (^6)]
(3)
(4)
Fstructure rates the strength of the image structure.
Because the intensity of the image structure varies both
between patients and also within a single dataset, a wide
interval of image intensities for high fissure probability is
defined. To estimate suitable values for  and ,  values
for five other datasets have been analyzed. From this
analysis it is revealed that  value lies between 20 and
80. Vessels which are high intensity component possess
high , whereas smaller vessels show value around 60.
is the Heaviside function that sets Fstructure value to
0 (i.e., a dark structure on a bright background), when≥
. Generally in medical application sensitivity is preferred
Copyright to IJIRSET
over specificity, hence  and  is chosen for the
distribution function of Fstructure. Therefore the voxels with
 value 30 to 70 is considered to be high fissure
probability voxels and voxels having  value greater than
100 are eliminated from segmentation. The Fsheet
differentiates between the sheet structure and other
structure like nodules, vessels etc., since the vessels and
nodules have larger  values compared with the fissure.
Similar to  and  values  is also determined from
the five other datasets which is not used for evaluation in
this paper. Thus voxels with the 2 value greater than 30
are assigned a probability 0, and eliminated from the
segmentation. The two features Fsheet and Fstructure should
be in the range [0, 1]. These two features are combined
using (5), to form the similarity measure Sfissure. The result
of fissure enhancement is that each voxel should have a
value in the range of [0, 1].
Sfissure = Fstructure*Fsheet
(5)
The result of fissure enhancement is used for the
segmentation of pulmonary fissure which is used for the
watershed lobe segmentation. A mask C, which describes
all the candidate voxels, is formed, such that each voxel
satisfies two constraints. One is the minimum fissure
similarity and the other demands an intensity value within
the specified range.
www.ijirset.com
M.R. Thansekhar and N. Balaji (Eds.): ICIET’14
1184
Automatic Segmentation of Pulmonary Lobe using Marker Based Watershed Algorithm
III. RESULTS AND DISCUSSION
C = [Sfissure > 0.1] ^ [Iv < (vessel – 2vessel)]
(6)
Where, the vessel and vessel are the mean and standard
deviation of the vessels, which are determined from the
histogram analysis of the segmented vessels. I v represents
intensity value of each pixel.
3) Bronchi: Lobes are separately supplied sub-trees of
bronchial tree. Hence distance to the bronchial tree is
suitable feature to detect lobar boundaries. In CT images,
the airway lumen is dark and separated from the other
parenchymal tissue by thin airway a wall structure that is
brighter. In CT images, the segmentation of airways is
challenging, since the parenchymal fissure often has same
Hounsfield Unit values similar to Lumen. Also the Partial
Volume Effect and noise obscure the airway walls. Two
pre-processing steps are applied to overcome these
problems. First one is Gaussian smoothening [9]. This
smoothening is applied to the image to reduce the noise,
even though this blurring increases the partial volume
related problems. The second is the, enhancement of the
bronchial tree [10]. The enhancement filtering is applied
to the blurred image. Due to partial volume effects and
additional Gaussian blurring the lumen of small airways
appear brighter than normal air. The importance of
bronchi enhancement filtering is to detect the voxels that
are surrounded by dense circular structure such as
bronchi. Finally to segment the bronchial tree, region
growing algorithm [11],[12], is used after applying
thresholding to the enhanced image. The region growing
is initialized by detecting the trachea pixels as the seed
point.
B. Watershed Lobe Segmentation
The various anatomical features that are extracted from
segmentation of fissures, blood vessel and bronchial tree
are combined into cost image [13]. This cost image is
formed by multiplicative method. To get good coverage
of the lobe area, the markers created should be distributed
throughout the lobes. The various sub trees belonging to
the various lobes and other lobar segments are estimated
to generate markers. For the generation of the markers
first a directed graph is generated from the bronchial tree
with trachea as the root [14]. The two sub trees with
maximum separation are separated into two different
branches. Similar to this separation, first the lung is
divided into left and right lung[15], and then the sub trees
are further divided into various lobes. Thus the labeled
airway is used to determine the marker position for the
watershed. Next step is to convert the labeled vessels,
bronchial tree and fissure as marker for the watershed
algorithm [7]. The border between lobes that is obtained
as a result of watershed segmentation is not smooth. This
is because of the local variations in the cost image. Two
majority filters are applied in row wise manner to smooth
the resultant segmented image.
Copyright to IJIRSET
The algorithm is implemented in MATLAB 2010a on
windows 7 environment. And it is tested with dataset
containing six images of different views. Fig.3(a), 3(b)
shows the images with incomplete fissures. The sample
image Fig.3 (a), considered here is of size (275×183),
similarly Fig.3(b) considered is of the size (75×111). As a
result of application of Hessian matrix, the Eigen values
are obtained. Using (3) and (4), Fstructure and Fsheet that
describe the characteristics of the image are obtained.
From these two parameters, the similarity index Sfissure is
determined from (5).
Fig.3(a),(b) images showing incomplete fissures
Table I. Fissure Enhancement Parameters
Parameters
Figure


Fstructure
Fsheet
Sfissure
3(a)
27.09
40.04
0.994
0.198
0.197
3(b)
24.95
45.66
0.999
0.741
0.74
Table 1 shows the enhancement parameters of the
Fig.3(a) and Fig.3(b), which is used for fissure
segmentation. In both the cases, λ3 value lies above 30,
hence the probability of being fissure is high. Similarly,
while considering λ2, if the value is between 0 and 30, the
identified feature is discriminated as fissure. Also the
parameter Sfissure lies in the range [0, 1].
www.ijirset.com
M.R. Thansekhar and N. Balaji (Eds.): ICIET’14
1185
Automatic Segmentation of Pulmonary Lobe using Marker Based Watershed Algorithm
the segmented image, which is shown in Fig.4(e2) and
Fig.4(f2).
Fig.4 Results obtained after segmentation
Cost image
(a)
Segmented
blood vessel
(a1), (a2)
Lobe
segmented
output
(b)
(b1), (b2)
Segmented
fissure
(c1), (c2)
Fig.5 Cost image formation and lobe segmentation
The three segmented images are combined into cost
image using fusion technique, which is the input for the
lobe segmentation process. Fig.5(a) shows the cost image
formed. Fig.5(b) shows the lobe segmented image of the
lung. The selection of markers is the first primitive step in
watershed algorithm, which helps to avoid over
segmentation. Subsequently watershed algorithm is
applied to obtain the segmented lung lobe. Fig.5(b) shows
the final segmented output.
(d1), (d2)
Segmented
bronchial
tree
(e1), (e2)
(f1), (f2)
The Fig.4(a1), Fig.4(b1) shows the coronal screen shot of
the lung image. This sample image, Fig 4(a1) is of the
size (228×215), similarly Fig.4(b1) is of size (265×190).
Both these images are rescaled to (255×255). After
rescaling thresholding is applied with a threshold value
130, which is determined from the histogram analysis.
The Fig.4(a2), Fig.4(b2), shows the segmented output of
blood vessels. The Fig.4(c1), Fig.4(d1) shows the axial
view of the lung image. The sample image Fig.4(c1)
considered is of size (275×183), similarly Fig.4(d1)
considered is of size (75×111). Based on the enhancement
parameters from Table 1, mask C is obtained from the (6),
which yields the segmented fissure output, which is
shown in Fig.4(c2), Fig.4(d2). Fig.4(e1), Fig.4(f1) shows
the 3-D view of lung image. Fig.4(e1) is of the size
(227×222) and Fig.4(f1) is of the size (194×259). These
images are first smoothened, and the blurred image is
enhanced. For enhanced image, region growing technique
is applied. This image is finally threshold by 90, to obtain
Copyright to IJIRSET
The proposed lung lobe segmentation method was
applied to dataset 1 images. The mean was calculated for
each lobar border. Table 2 shows the mean calculation
and comparison with the existing methods by Van
Rikxoort et al which show automatic segmentation robust
against incomplete fissures [4] and Kuhnigk et al which
proposed new tools for computer assistance in thoracic
CT.
Table II. Results Showing Mean of Various Methodologies
Methods
Mean
Fissure alone is considered for
Segmentation of Lobe
2.78
Cost Image using Euclidean
Distance and ROI
1.08
Proposed Method
0.86
From the Table II, it is clear that the proposed method
shows better performance than the existing methods like
Rikxoort and Kuhnigk.
The proposed system is robust against incomplete
fissure, but in cases of missing fissure the lobar boundary
www.ijirset.com
M.R. Thansekhar and N. Balaji (Eds.): ICIET’14
1186
Automatic Segmentation of Pulmonary Lobe using Marker Based Watershed Algorithm
is not exactly defined. The accuracy is determined from
(7)
REFERENCES
[1]
Accuracy = (TP+ TN) / (P + N)
(7)
Where TP (True positive) and TN (True negative) are
determined from the confusion matrix and P (Positive)
and N (Negative) are estimated from the ground truth.
Table III shows the accuracy calculation of the proposed
methodology. A set of twenty images with incomplete
fissure and missing fissure is considered for the accuracy
calculation. Out of all 100 lobes in twenty images 84 was
correctly segmented.
[2]
[3]
[4]
Table III. Accuracy Calculation
[5]
Number of
Images
Taken for
Analysis
Total
Number of
Lobes
Number of
Lobes Correctly
Segmented
Accuracy
(%)
25
125
105
84%
[6]
[7]
Table IV. Run Time Analysis
[8]
Schemes
Run time/ lung
Manual
75 min
[9]
Implicit surface fitting
25-30 min
Atlas based method
15 min
Proposed method
10 min
[10]
[11]
The run time measures are tabulated here. Table IV
shows the comparison of the proposed algorithm and the
existing algorithms like atlas based, implicit surface
fitting [10]. The run time is considerably reduced when
compared to the existing methods.
[12]
[13]
IV. CONCLUSION
This paper presents, an algorithm for automatic
segmentation of lung lobe, which is robust against
incomplete fissures. Fissures or lobar boundaries play a
major role in diagnosing the lung disease. Nowadays, lung
cancer is popularly seen among the people. Due to lung
cancer the lung gets severely damaged. In such cases of
severely affected lung, the fissures are incomplete. This
algorithm performs lung lobe segmentation which
overcomes the defect of exciting by including features
like pulmonary vessels and bronchial tree.
Copyright to IJIRSET
[14]
[15]
E. van Rikxoort, B. van Ginneken, M. Klik, and M. Prokop,
“Super- vised enhancement filters: Application to fissure detection
in chest CT scans,” IEEE Trans. Med. Imag., vol. 27, no. 1, pp. 1–
10, Jan. 2008.
J. Pu, J. Leader, B. Zheng, F. Knollmann, C. Fuhrman, F. Sciurba,
and D.Gur, “A computational geometry approach to automated
pulmonary fissure segmentation in CT examinations,” IEEE Trans.
Med. Imag., vol. 28, no. 5, pp. 710–719, May 2009
E. M. van Rikxoort, B. de Hoop, S. van de Vorst, M. Prokop, and
B. van Ginneken, “Automatic segmentation of pulmonary
segments from volumetric chest CT scans,” IEEE Trans. Med.
Imag., vol. 28, no. 4, pp. 621–630, Apr. 2009.
E. van Rikxoort, M. Prokop, B. de Hoop, M. Viergever, J. Pluim,
and B.van Ginneken,“Automatic segmentation of pulmonary lobes
robust against incomplete fissures,” IEEE Trans. Med. Imag., vol.
29, no. 6, pp. 1286–1296, Jun. 2010
S. Ukil and J. M. Reinhardt, “Anatomy-guided lung lobe
segmentation in x-ray CT images,” IEEE Trans. Med. Imag., vol.
28, no. 2, pp. 202–214, Feb. 2009.
B. Lassen, J.-M. Kuhnigk, O. Friman, S. Krass, and H.-O. Peitgen,
“Automatic segmentation of lung lobes in CT images based on
fissures, vessels, and bronchi,” Proc. Int. Symp. Biomed. Imag.,
pp. 560–563, 2010.
Bianca Lassen, Eva M. van Rikxoort, Michael Schmidt, Sjoerd
Kerkstra,
Bram
van
Ginneken,
and
Jan-Martin
Kuhnigk,”Automatic Segmentation of the Pulmonary Lobes From
Chest CT Scans Based on Fissures, Vessels, and Bronchi” IEEE
Trans. Med. Imag., Vol. 32, no. 2, Feb 2013 .
Jiri hladuvka, Andreas Konig, Edward Groller, “Exploiting
Eigenvalues of the Hessian Matrix for Volume Decimation”,
Institute of omputer Graphics and Algorihms, Vienna University of
Technology, Karlsplatz 13/186, A-1040 Wien, Austria.
P. Spureka, A. Chaikouskayaa, J. Tabora, E. Zaja¸cb, “A local
Gaussian filter and adaptive morphology as tools for completing
partially discontinuous curves”, arXiv:1311.7430v1 [cs.GR] 28
Nov 2013.
S. Arastehfar1, A. A. Pouyan2*, A. Jalalian1, “An enhanced
median filter for removing noise from MR images”, Journal of AI
and Data Mining , Vol.1, No.1, 2013, 13-17.
Huiyan Jiang,1,2 Baochun He,1 Di Fang,1 Zhiyuan Ma,1
Benqiang Yang,3 and Libo Zhang3, “A Region Growing Vessel
Segmentation Algorithm Based on Spectrum Information”,
Computational and Mathematical Methods in Medicine Volume
2013, Article ID 743870, 9 pages.
H.P. Narkhede, “Review of Image Segmentation Techniques”,
International Journal of Science and Modern Engineering
(IJISME) ISSN: 2319-6386, Volume-1, Issue-8, July 2013.
Rohan Ashok Mandhare, Pragati Upadhyay, Sudha Gupta, “PixelLevel Image Fusion Using Brovey Transforme And Wavelet
Transform”, International Journal of Advanced Research in
Electrical, Electronics and Instrumentation Engineering , Vol. 2,
Issue 6, June 2013
B.Sridhar, Dr.K.V.V.S.Reddy, Dr.A.M.Prasad, “Automated
Medical image segmentation for detection of abnormal masses
using Watershed transform and Markov random fields”, ACEEE
Int. J. on Signal and Image Processing , Vol. 4, No. 3, Sept 2013.
D. Selle, B. Preim, A. Schenk, and H.-O. Peitgen, “Analysis of
vasculature for liver surgical planning,” IEEE Trans. Med. Imag.,
vol. 21, no. 11, pp. 1344–1357, Nov. 2002.
www.ijirset.com
M.R. Thansekhar and N. Balaji (Eds.): ICIET’14
1187